sai-prasanna / dreamer_augment

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Dreamer Augment

Dreamer 2

We build our work upon the Dreamerv2 pytorch implementation.

Environment

In a conda or virtual env, install the requirements.

pip install -r requirements.txt

Training

The dreamer-torch/config.yaml has various configurations that can be selected using --config argument and every individual keys can be overridden using --key.

Note:

  1. Setting batch size to 156 is important for sample efficiency.
  2. You can disable weights and biases logging by passing --wandb False.

Few examples are below.

Dreamerv2 Baseline

cd dreamer-torch
python dreamer.py --configs defaults dmc --logdir fs_baseline --seed 42 --task dmc_finger_spin --wandb_name fs_baseline

Dreamerv2 CURL

cd dreamer-torch
python dreamer.py --configs defaults dmc --logdir ch_curl --size 84,84 --seed 1337 --task dmc_cheetah_run --augment True --augment_random_crop True --augment_pad 0 --augment_crop_size 64,64 --augment_strong False --augment_consistent True --wandb_name ch_curl --curl True --batch_size 150 --seed 1337 --batch_size 156

Evaluation

Use dreamer-torch/dreamer_eval.py for running evaluation.

Dreamer 1 Rllib

Our prior experiments used rllib's dreamer1 implementation, which we modified to include some elements from dreamer2. For running them, check the dreamer/train.py. The configurations can be run together with grid search.

About


Languages

Language:Python 100.0%